Agent Relations @FactoryAI I code often and sleep occasionally. I like systems. Agents are graphs. Harness is everything. Views are mine. #2 fan of Corvus Corax
Night City
Joined November 2023
- Tweets5.6K
- Following3.2K
- Followers610
- Likes55.4K
Ainesh Chatterjee retweeted
A ‘hello world’ milestone for Project Suncatcher 🚀
After launching our prototype satellite a few weeks ago, we’ve now successfully sent a command up to orbit, booted up the TPU, ran the prompt, and received our first complete response back on earth - a haiku!
Ainesh Chatterjee retweeted
1/4
Meet Qwenputer: Qwen3.5-4B with a full RISC-V CPU inside its weights, using the model's own Gated DeltaNet layers as RAM.
It runs C and MicroPython in its forward passes, no external runtime.
On DeepMind math: base 55–90%, base + Python tool 95–100%, Qwenputer 97.5–100%.
A lot of late nights, wouldn’t want to have done it with anyone else!
incredibly excited to announce that droid is now v1.0 - we've modernized just about every part of our cli, including a full rewrite to the agentic engine that powers it.
reflecting on the last few months leading this project, it has been awesome to see how heavily the team has embodied factory's operating principle of "build together," which is timely considering the doomer phase our industry is going through.
"software factory" is a term everyone is throwing around these days but it often gives the vibe that humans are no longer in the loop. for factory, this couldn't be further from the truth. our agent factory is human-first: @ain3sh, @wenemilyy, @agent_crunchwrap and i went through every PR droid proposed, threw out the changes that overfit to a single model, and pushed droid to keep going. we collaborated on sessions together, experimented with hundreds of approaches, and fought the slop we left behind via huge refactors that we only had confidence in due to the extensive e2e tests we invested in half a year ago.
getting to this point wasn't easy. it involved challenging our priors on how agents should work, deviating from old patterns, and setting innovative goals for our harness to hillclimb itself towards.
attached is the video of the process that got us here - the result is the world's best agent.
droid is back in first place, and we're here to stay.
Ainesh Chatterjee retweeted
is your window manager too rectilinear? do you wish it had more curves? are you tired of the same old euclidean metric, day in and day out?
do i have the tool for you!
introducing, Horosphere WM: the first wayland compositor to use hyperbolic geometry to organize your workspace
Ainesh Chatterjee retweeted
Anthropic rewrote its Messages API in Rust. Zero Rust devs on the team.
Claude wrote the code. Humans wrote the plan and checked old vs new matched on real traffic. CPU dropped 40-50%.
Pretty sure "never do the big rewrite" is dead advice.
@katelyn_lesse on @latentspacepod
Ainesh Chatterjee retweeted
Really cool to see this on the feed because I work on sandboxes at Factory and this ephemeral vs. long-lived shape is literally what we've been stewing on for the last year. Earlier this year, we decided to reshape our remote sandboxes into the latter for many of the reasons you described: factory.com/news/droid-compu…
However, we actually brought back "ephemeral" per-session machines recently factory.com/news/droids-now-… and ever since we rolled out Environments internally, the % of internal usage we have on remote vs local has gone up significantly. "Ephemeral" is in quotes here because these are just persistent Droid Computers under the hood and still retain local state after sleeping and waking up, so when I do do work that requires a local DB for example, I can just go back to the session where I originally seeded the DB, or I can still just create a Droid Computer explicitly for this purpose. But 99% of my work doesn't require that, and now I can fan out all of that work to many different agents without having them bottlenecked on CPU/memory/disk allocated to my single devbox or having to do worktree gymnastics. And now that agents are so capable, instead of just relying on local state persisting to keep my dev setup updated to my tastes, I can have the agent keep updating the Environment definition so that all new sandboxes get new changes I made in my previous sessions.
Also happy to say we've thought through a lot of the other downsides mentioned here
> I want my agent to read my datadog logs to help me debug an issue, how does that work?
We have an Axiom connector that works from both local and remote environments factory.com/product/connecto…
> I want my agent to be able to hit internal services on my Tailnet
> Another problem is: there's also security concerns about using 3P sandboxes
We have a self-hosted sandboxes product that lets you spin up sandboxes from Factory in your own VPC that can connect to your internal services factory.com/news/opencompute
> I want to build & smoke-test & deploy docker containers, and ephemeral agents run into dind issues
dind works in both our managed and OpenCompute sandboxes. Personally have done a lot of meta work on sandboxes in our sandboxes
> I want my agent to help me zip up some files, inspect them for a while + put them somewhere else. (likely AWS access bottlenecked?)
> Devbox + tunnel for localhost + agent docs on how to use the tunnel is really sweet
File upload/download and port-forwarding are first-party features on Droid Computers
Anyways obviously biased towards the product I've been working on, but this is just to say that ephemeral vs. persistent isn't exactly a strict dichotomy -- it's more about single-session vs multi-session scope, and I wouldn't rule out session-scoped sandboxes because they can be done well. Clearly there's a place for both shapes. And also we're cooking up some good stuff that takes advantage of our ability to do both :)
We’re planning a lot more in terms of the underlying architecture for the CLI, etc. for @droid that powers all of the Factory.
But this was a crucial foundation to lay to make those more ambitious things practical.
So hyped, and very grateful, to get to work on these things with @agent_wrap 🥹
Ainesh Chatterjee retweeted
I've been thinking about this a bunch lately. Codebases have become organic, perhaps they always were, and we require new abstractions to do excellent engineering.
Cybernetics is back. What's old is new again, and this is the current best way to think about it. For all their flaws, Anthropic has the *exact* right idea with "SuperFans". I mean, I'm sort of in awe of how good that particular blog post was. There's this commitment to excellence that I greatly admire, and a level of AGI-pilled reasoning that meets the moment.
In my opinion, you should work with models you trust and then give them the autonomy to own part of your codebase. This doesn't mean they don't get audited, or manually checked, or interviewed - it just means that each component gets a type of guy who is obsessed with that component.
If you think about this from a higher level of abstraction, what you're really doing is enabling *zones of dynamic attention* where the models are using their fluid intelligence to design a component and then justify that component in the context of all the others. If a new one gets built that may cause issues, that model works with the model for the other component to ensure mutual compatibility. It's almost like the components themselves have come to life and are trying to figure out how best they might fit together.
The old ways of engineering are over. Codebases that are built like this will be far more competitive almost inevitably as the skill of the model surpasses that of the average programmer and sheer parallelization wins the day. Obsession over PRs, dedication to reading every line, these are vestigial neuroses that must be abandoned if we are to succeed well into the future.
Operational excellence without constant human oversight is possible with good enough models and clever enough scaffolds. The hardest part is almost always going to be turning it on and managing the first few problems. These systems will probably never be perfect, but they are the path forward.
we will need to rebuild and rewrite much of software to adapt to a new, faster pace.
code used to arrive one drop at a time. soon, if not already, code will be written as if it were a continuous stream, like a violent river twisting and turning as it pleases.
this is not a new thing: big tech cos have had to deal with this with the enormous scale of their engineering teams. but now ai has brought an incredibly elastic and capable workforce wielded by every engineer. this is your problem now, and the only way to adapt is to evolve your systems so agents - even and especially the ones with little context about the right way of doing things - can thrive. make the easy thing the right thing.
your codebase is becoming organic architecture. agent contributions grow around the constraints and guardrails you’ve designed. good software engineering matters even more now! and no, the continuous stream of code is not something you have to watch helplessly. be more ambitious and design systems, frameworks, and languages to control the chaos.
the time for mourning coding is long over
I still remember kicking off a Jules task before heading into a lecture, feeling like it was a new world 🥹
a pure attempt, rip
When we launched Jules, it was an experiment to learn how asynchronous coding agents can help developers ship faster. Seeing what you built taught us so much about where agentic development is headed, directly shaping several products across @Google.
We’re working on bringing our learnings from Jules into @antigravity so you get the best of both worlds: autonomous remote background task execution and an agent-first development environment.
As we focus on this unified experience, Jules will be winding down.
Whether you used Jules to crush bugs, write tests, or launch features— thank you for shaping the future of agentic coding! 💙🦑
Ainesh Chatterjee retweeted
Last night an AI agent performed brain surgery on a mouse.
We gave it 19 gene therapy vectors in a spreadsheet, a mouse model of epilepsy, access to a motorize surgery rig with cameras and one goal: reduce seizures.
It picked the drug. It picked the brain region. It derived the coordinates from three papers. Then it drove the stereotaxic arms into the brain and injected 250 nanolitres while watching the entry site through the rig camera for bleeding.
A licensed vet gated every irreversible step and held a physical kill switch to stop the AI agent.
Thread on what it decided, what it saw, and the moment it refused to let me override it.
[Writing this on the spot out of excitement, but I've been 18hrs in front of my laptop coding and coordinating this, so keeping it brief and descriptive, more cohesive thoughts on this soon...]
Harness matters.
Since the days of TB2, other offerings in the space are now much more considerate with their harness' design, but experience, relentlessness, grit, and ofc principled research, works wonders!
Ainesh Chatterjee retweeted
software is changing. I think every app will be malleable by its user, from how it looks to how it works. herdr-app is my bet on that.
and if you're building custom agent orchestration, herdr-core is really cool. give it a shot.
Ainesh Chatterjee retweeted
*ehm ehm* BIG ANNOUNCEMENTS DAY 📣
two new things cooking in herdr labs 🐏🧪
a brand new herdr-core, and a desktop app built on it. open source.
first, the app: software that fits you. every part of it is a file your agent can change. you just ask.
Ainesh Chatterjee retweeted
i think this is real and a big deal. as karpathy said, the ugliest part of the transformer is the tokenizer
I've always believed that intelligence should come from simple, general building blocks that learn their own abstractions directly from raw data; in other words, the bitter lesson. I think we're finally seeing a real step towards AI that can learn from all data.
Language modeling on raw bytes is a microcosm of this ideal, and to me it's a central problem in AI: if we can't build LLMs that learn end-to-end from bytes, how will we handle the rest of the world's data that we understand even less?
Last year we introduced hierarchical networks (H-Nets) as a first step. Our new version finds up to 3-5x efficiency gains in controlled scaling law experiments, and at larger scale matches an open-source model trained on 11T tokens with ~10x less data and compute. The posts below walk through what we found and why we think it works.